Consider casting to float32 by default in TableVectorizer
GaelVaroquaux opened this issue · comments
Gael Varoquaux commented
Problem Description
Using float64 instead of float32 typically incurs compute and memory loads, and users do not have this in mind.
Feature Description
We should add an option to the TableVectorizer to output float32. We should consider whether this is the default.
Alternative Solutions
N/A
Additional Context
N/A
Jérôme Dockès commented
that also applies (maybe even more) to encoders, for example MinHash outputs float64
Gael Varoquaux commented
that also applies (maybe even more) to encoders, for example MinHash outputs float64
Absolutely! Thanks for raising this. Maybe we should start there
Théo Jolivet commented
Closing because this has been addressed in the postprocessing step of the TableVectorizer
in #902